Partial Paper Mask Attack Dataset
3k+ videos of partial facial fragment attacks for anti-spoofing and liveness detection
Check samples on Kaggle
Dataset summary
|
Parameter
|
Value
|
|---|---|
|
Volume
|
3,000+ videos from 50+ unique IDs
|
|
Coverage
|
Partial paper mask presentation attacks for face anti-spoofing
|
|
Demographics
|
Adults, mixed gender, multi-ethnic
|
|
Devices
|
iOS and Android phones (dual-device capture)
|
|
Conditions
|
Indoor, varied lighting
|
Introduction
The Partial Paper Mask Attack Dataset is a collection of 3,000 videos featuring partial paper mask attacks – printed fragments of another person’s facial features overlaid on the attacker’s real face. Unlike full-face cutout print attacks where the entire photo is used, these attacks target specific facial regions: eyes, mouth, nose area, and their combinations, while the rest of the attacker’s real face remains exposed. This makes partial paper mask attacks significantly harder to detect: real skin texture, natural blinking, and live facial movement co-exist with the printed fragment, challenging both texture-based and motion-based anti-spoofing systems
Dataset Features
- Fragment types: Multiple printed facial fragment configurations including eye region, nose region, mouth region, and combined upper- or lower-face overlays
- Per-participant structure: Approximately 60 videos per participant on average, distributed across fragment types, devices, and capture conditions, providing the within-identity variation needed to train models that generalize beyond surface texture cues
- Capture protocol: Every video follows a standardized active-liveness sequence – zoom-in, zoom-out, head turn left and right, and natural blinking, mirroring the challenges issued by production liveness systems during real verification flows
- Skin-tone matching: Paper fragments are printed on color-calibrated stock and matched to each participant’s natural skin tone prior to each session. This removes the obvious tone-mismatch artifacts that simple color-based and texture-based detectors rely on, forcing models to learn deeper liveness signals
- Lighting conditions: Captures span three lighting categories – low-light indoor, bright indoor, mixed indoor-window, to expose models to the full range of real-world deployment conditions
- Video duration: ∼10 seconds per clip at native device frame rate
How It Differs From Cutout Print Attacks
|
|
Cutout Print Attack
|
Partial Paper Mask Attack
|
|---|---|---|
|
What is used
|
Full-face photo cutout
|
Printed fragment of a specific facial region
|
|
Real face visible
|
No, photo covers entire face
|
Yes, attacker's real skin, hair, and features are exposed
|
|
Liveness features
|
None, fully static print
|
Real blinking, movement, skin texture alongside the fragment
|
|
Detection difficulty
|
Standard for PAD systems
|
Harder, mix of real and fake signals
|
Real-World Validation: Open Liveness Model Stress Test
To demonstrate the practical value of this dataset, we tested its samples against Doubango’s open-source face liveness model – a publicly available liveness detection SDK used as a reference implementation by anti-spoofing researchers and developers.
Key result: Doubango’s liveness classifier rated the partial paper mask attack below as 88.69% genuine – failing to detect the spoof at the liveness layer. The attack was only flagged by Doubango’s secondary injection-detection module, not by liveness scoring itself
Use cases and applications
Face Anti-Spoofing & PAD: Train presentation attack detection models to identify partial paper overlay attacks – a blind spot for systems trained only on full-face prints, replay attacks, or 3D masks
Liveness Detection: Improve liveness detection robustness by exposing models to attacks where real facial movement co-exists with printed fragments, challenging motion-based and texture-based detectors simultaneously
- iBeta Certification Preparation: Test your liveness system against realistic 2D partial attacks before submitting to iBeta Level 1 or Level 2 certification
iBeta Success Stories
Datasets like this contributed to 21% of companies that passed iBeta certification in 2025 – all Axon Labs clients
From Failed Attempt to Success: UK Fintech Passed iBeta 2
UK fintech invested in datasets from another vendor for iBeta 2 and failed. The gap analysis revealed critical missing attack vectors. They needed a solution fast
Double Success: E-KYC Startup Achieves iBeta 1 & 2
A US-based e-KYC startup needed to prove their facial recognition technology met the highest industry security standards. Their ambitious goal: pass iBeta 1 and 2 Certification
Technology company from Vietnam: iBeta Level 2 success
A Vietnam-based AI/Big Data firm coached by Axon Labs passed iBeta PAD Level 2 on the first try with 0% successful spoofs; the solution claims 99.9% face-recognition accuracy
Source and collection methodology
All videos were recorded in-house by the Axon Labs team following a controlled recording protocol. Facial fragments are printed on high-quality color paper, skin-tone matched, and proportionally sized to the participant’s face. Videos include active liveness features: zoom-in/out, natural head movements, and blinking. Recordings were captured across diverse backgrounds and lighting conditions
Need More Data?
This dataset is a ready-made sample. We offer custom data collection for paper mask attacks tailored to your requirements, including larger participant pools, additional devices, specific demographic distributions, and custom mask configurations
Download information
A sample version of this dataset is available on Kaggle. Leave a request for additional samples in the form below
Have a question?
All data is collected in-house by the Axon Labs team. All information is verified by our specialists
Once your enquiry has been sent, we will contact you to discuss the details and complete the necessary paperwork. The timing of receiving the dataset depends on the specific request and additional requirements
Our unique selling point is to provide legally clean datasets to our customers. We obtain the consent from all the participants to use their data for AI model development. We are able to provide comprehensive reporting on the licensing, data collection and privacy compliance of our datasets
The price depends on your specific requirements. Please submit a request to receive a free consultation
The Cutout Print Attack dataset uses full-face photo cutouts — the entire printed face replaces the attacker's face. This dataset uses partial facial fragments: only specific regions (eyes, mouth, nose) are printed and overlaid, while the attacker's real face remains visible around the fragment. This creates a fundamentally different challenge for detection systems
Contact us
Tell us about yourself, and get access to free samples of the dataset
Didn't find what you were looking for?
Our collection includes many datasets for various requests
iBeta Level 1 Dataset
– 35,000+ videos
– 85+ participants
– zoom in and
zoom out
iBeta Level 2 Dataset
– 25 000+ videos
– 3D masks
– iBeta Level 2
iBeta Level 3 Dataset
– 10,000+ videos
– 12 Unique Masks
– iBeta Level 3
Display Replay Dataset for Liveness Detection
– 9,000+ videos
– 6,500+ participants
– Balanced mix of genders and ethnicities
